Researches On Image Registration Based On Local Feature Information

The technology of Image registration has been widely applied to many fields which have close correlation with people’s life, such as terminal guidance of the military application, medical image processing, Three-Dimensional modeling, image mosaic, etc. In the application of terminal guidance, how to recognize the target efficiently and accurately in complicated scenes is a main problem. A target object may have significant difference in texture, structure and location in different images obtained in the result of the difference of the sensor, filming moment and perspective. It is difficult to get a reasonable result when we use the methods based on the texture information of region to deal with this kind of situation. The methods based on local invariant features have obvious advantages in time consumption and matching result in contrast with methods based on the texture information of region. Therefore, this paper makes a study on the technology of image registration based on the analysis of local feature information.This paper analyzes the general process and the technological means in the application of terminal guidance. Two salient features have been chosen based on the different application requirements to do the matching process after analyzing the difference between the images obtained by different perspectives and shooting conditions:One is point feature, the other is line feature.We mainly focus on the improvements of matching process when handling with the point feature. The matching accuracy is not guaranteed since there are many similar areas in global scope and this may lead to the mismatch because of the point feature descriptor is a local image texture abstract description. In order to solve this problem, we proposed a progressive matching algorithm based on the point feature. Firstly, we build up the initial matching correspondence after feature point detection process and filter the result by RANSAC. Secondly, the initial matching correspondence is been used as prior information in the progressive matching stage. Finally, we establish the final matching correspondence by least squares estimation. Experimental results show that the results of proposed method have obvious advantage than the other methods on matching accuracy and the correct number of matching points.We regard contour feature as the main object in this paper. A contour may have different changes on different perspective. It is difficult to get reasonable results when we simply do some similarity analysis on contours. In order to solve this problem, we proposed a contour matching method which is based on the curvature scale space. We eliminate the influence of affine transformation by PCA whitening and ICA. Then we find some key points which have affine invariant on the contour by taking advantage of the curvature scale space image. These points are used to match instead of matching contours directly. Experimental results show that the proposed method has reasonable results when handling the contour matching process with affine transformation.